Underinvestment in Profitable Technologies when

Experimenation is Costly:

Evidence from a Migration Experiment In Bangladesh

Gharad Bryan, Shyamal Chowdhury and A. Mushfiq Mobarak

April 1, 2011

1 Introduction

The causes and consequences of internal migration have received comparatively little at-tention in the economics and policy literatures, even though it is much more common thanthe heavily-studied issue of international migration (Borjas 1999, Rosenzweig 2005, Yang2008, Hanson 2009). There were 240 times as many internal migrants in China in 2001 asthere were international migrants (Ping 2003), and 4.3 million people migrated internallyin the 5 years leading up to the 1999 Vietnam census compared to only 300,000 inter-national migrants (Ahn et al, 2003). The development effects of internal migration arearguably more profound: the majority of households in rural Rajasthan migrate seasonallyin search of employment, and use this as the primary vehicle for diversifying sources ofincome (Banerjee and Duflo 2006).

This paper studies the causes and consequences of internal seasonal migration in north-western Bangladesh, in a region where over 5 million people below the poverty line mustcope with a pre-harvest seasonal famine, known locally as Monga, almost every year (TheDaily Star 2010). This seasonal famine is emblematic of widespread pre-harvest “lean” or“hungry” seasons throughout South Asia and Sub-Saharan Africa documented in a num-ber of studies,1 where imperfect consumption smoothing forces households below povertyduring parts of the year. Inspired by the observation that nearby urban areas offer bet-ter wage and employment opportunities during the lean season, we provide small grantand loan incentives (of $8.50 or 600 Taka) in 100 study villages to encourage people toseasonally migrate out in search of employment. The random assignment of incentives

1Seasonal poverty or hungry seasons have been documented in Ethiopia (Dercon and Krishnan 2000,who show that poverty and malnourishment increase 27% during the lean season), Malawi (Brune et al2010), Madagascar (Dostie et al 2002, who estimate that 1 million people fall below poverty before the riceharvest), Kenya (Swift 1989, who distinguishes between years that people died versus years of less severeshortage), Thailand (Paxson 1993), India (Chaudhuri and Paxson 2002) and China (Jalan and Ravallion1999).

1

allows us to generate among the first experimental estimates of the effects of migration,and internal migration in particular.Estimating the returns to migration is the subject ofa very large literature, but one that has been hampered by difficult selection issues (Akee2010, Grogger and Hanson 2010, McKenzie et al 2010). The internal seasonal migration weencourage is a commonly used means to cope with seasonal poverty and natural disastersacross Asia, Africa and Latin America,2 and our estimates of the effects of migration thushave substantial external relevance.

We estimate the causal effects of being induced to migrate away during the lean seasonto be very large.These estimates add to an emerging literature that documents very highrates of return to small capital investments in developing countries (Udry and Anagol200x, De Mel, McKenzie and Woodruff 200x and 200x and Duflo, Kremer and Robinson2009), and bolsters the case made by Clemens and Pritchett (20xx), Rosenzweig and xx(200x) and Gibson and McKenzie (2010) that offering migration opportunities improveswelfare by much more than any other development intervention in health, education oragriculture that has been studied. Migration induced by our intervention increases foodand non-food consumption of the migrants family members remaining at the origin by30-35%, and improves their caloric intake by 700 calories per person per day. On an initialinvestment of about $6-$8 (the average round-trip cost to a destination), migrants earn$110 on average during the lean season and save about half of that, which are suggestive ofa very high rate of return on investment.Most strikingly, these induced migrants continueto re-migrate at a higher rate compared to a control group a year later, even after theinducement for migration is removed.These large positive returns, consumption effects andpreferences revealed by the voluntary re-migration beg one very important question: Whydidnt these people already engage in such a highly profitable investment?

To understand this puzzling behavior, we propose a model in which experimentingwith a new technology or behavior such as migrating to a new destination - is risky, andrational households do not migrate in the face of uncertainty about their prospects at thedestination even when they expect monetary returns to the activity to be positive. Thisresults in a poverty trap in which households do not migrate out of fear of an unlikely,but devastatingly negative outcome where they undertake the cost of moving but returnhungry after not finding employment during a period when their family is under the threatof famine. Inducing the inaugural migration by insuring against this devastating outcomewhich our grant or loan with implied limited liability managed to do can lead to long-run benefits where households either learn how well their skills fare at the destination, orimprove their future prospects by allowing employers to learn about them. This simplemodel can explain the high take-up rate for the intervention, the large positive incomeand consumption effects, and the greater re-migration in a future period among treatment

2Prior attempts use controls for observables (Adams 1998), selection correction methods (Barhamand Boucher 1998; Acosta et al 2007), instrumental variables methods (BenYishay 2010, McKenzie andRapoport 2007, Brown and Leeves 2007, Yang 2008) and natural policy experiments (Clemens 2010, Gibsonet al 2010) to answer this question.

2

households even after the inducement is removed.The proposed model also makes four other predictions regarding migration behavior.

First, households that are close to subsistence – on whom experimentation with the newtechnology imposes the biggest risk – should be less likely to migrate and more likelyto benefit from our intervention. Second, households that do not know someone at thedestination and therefore do no have the contacts necessary for successful migration shouldbe less likely to migrate and more likely to respond to our incentive. Third, because thepoverty trap is generated by the possibility of a large negative return and that possibilitycan be effectively mitigated through a loan that does not have to be paid off if the migrationepisode was not successful, we predict that cash and credit incentives should have roughlythe same effect on the migration rate. Fourth, the possibility of learning implies that thosehouseholds that are most successful should be those most likely to repeat migrate and thatrepeat migration will be to the same location as prior migration. We find support for all ofthese conclusions in our data, the last one tested by using additional exogenous variationin migration location induced by conditionalities placed on migration in our experiment.

Households reluctance to engage in risky or costly experimentation can provide insightinto a number of other important puzzles in growth and development. Green revolutiontechnologies led to dramatic increases in agricultural productivity in South Asia (Evensonand Gollin 2003), but take-up and diffusion of the new technologies was surprisingly slow,partly due to low levels of experimentation and the resultant slow learning (Munshi 2004).More directly, research has linked the inability to experiment due to uninsured risk to abias towards low risk low-return technologies that stunt long-run growth (Yesuf et al 20xx),and to reduced investments in agricultural inputs and technologies such as new high-yieldvariety seeds and fertilizer (Rosenzweig and Wolpin 1993a, Dercon and Christiansen 2009).Aversions to experimentation can also hinder entrepreneurship and business start-ups andgrowth (Hausmann and Rodrik 2003; Fischer 2009), and can force sales of productiveinvestment assets in order to smooth consumption (Rosenzweig and Wolpin 1993b). Wecontribute to this literature by using experimental variation in data to build evidence onthe presence of a poverty trap induced by inefficiently low levels of experimentation.

Our model and experimental results shed light on a fundamental puzzle in developmenteconomics, which is that adoption rates of highly efficacious technologies with the potentialto address important development challenges - ranging from tropical diseases to low agri-cultural productivity to low rates of savings and investment - have remained surprisinglylow (Miller and Mobarak 2010, Null et al 2010). If adoption is risky (e.g. due to risk ofcrop failure, or uncertainty about durability of a new stove or water purifying technology)then giving households the opportunity to experiment with the new technology by insuringagainst failure may be an effective marketing strategy.

Finally, from a narrow policy perspective, our experiments uncover a cost-effective re-sponse to the widespread famines that afflict the 10 million people residing in Rangpurregion of Bangladesh with disturbing regularity. Such lean seasons are not phenomenaunique to northwestern Bangladesh: there is a predictable pre-harvest hungry season every

3

year in Malawi, Ethiopia, Madagascar and other countries, forcing millions to succumb toseasonal poverty. The solution we implement is inexpensive (the vast majority repay the$8.50 inducement when it is offered as a loan), it confers long-run benefits even when of-fered as a one-off, and it is more sustainable than subsidizing food purchases on an ongoingbasis, which is the major anti-famine policy tool currently employed by the Bangladeshgovernment (Khandker 2010; Government of Bangladesh 2005, journalistic cite). Our in-tervention mitigates the spatial mismatch between where people are based, and where thejobs and excess demand for labor are during the pre-harvest months. This approach maybe of relevance to other countries that face geographic concentrations of poverty, such asnorthern Nigeria, eastern islands of Indonesia, northeast India, southeast Mexico, and in-land southwest China (Jalan and Ravallion 2002). Of course one has to be careful aboutgeneral equilibrium effects at the destination when such a program is scaled up, but in-stances of excess labor demand during particular seasons is not altogether uncommon evenin developing countries. Our findings suggest that providing credit to enable householdsto search for jobs, and aid spatial and seasonal matching between potential employersand employees may be a useful way to augment the microcredit concept currently morenarrowly focused on creating new entrepreneurs and new businesses.

The next two sections describe the context and the design of our interventions. Wepresent results on program take-up and the effects of migration in Section 4. These findingsmotivate the risky experimentation model in Section 5. We present statistical tests ofvarious implications of the model in section 6, discuss alternative explanations of the datain Section 7 and offer conclusions and policy advice in section 8.

2 The Context: Northwestern Bangladesh and the MongaFamine

Our experiments are conducted in 100 villages in two districts (Kurigram and Lalmonirhat)in the seasonal-famine prone Rangpur region of north-western Bangladesh. The Rangpurregion is home to roughly 7% of the countrys population, or 9.6 million people. 57% of theregions population (or 5.3 million people) live below the poverty line.3 In addition to thelevel of poverty, the Rangpur region experiences more pronounced seasonality in incomeand consumption, with incomes decreasing by 50-60% and expenditures on food droppingby 10-25% during the post-planting and pre-harvest season (September-November) for themain Aman rice crop (Khandker 2010). As Figure 1 indicates, the price of rice also spikesduring this season, particularly in Rangpur, and thus kilos of rice consumed drops 22%

3Extreme poverty rates (defined as individuals who cannot meet the 2100 calorie per day food intakeeven if they spend their entire incomes on food purchases only) were 25 percent nationwide, but 43 percentin the Rangpur districts. Poverty figures are based on Bangladesh Bureau of Statistics (BBS) HouseholdIncome and expenditure survey 2005 (HIES 2005), and population figures are based on projections fromthe 2001 Census data.

4

even as households shift monetary expenditures towards food (Figure 2) while waitingfor the Aman rice harvest. The lack of job opportunities and low wages during the pre-harvest season and the coincident increase in grain prices combines to create a situation ofseasonal deprivation and famine (Mahmud and Khandker 20xx, Sen 1981)4. The famineoccurs with disturbing regularity and thus has a name Monga. It has been described as aroutine crisis (Rahman 1995), and its effects on hunger and starvation widely chronicled inthe local media. Agricultural wages in the Rangpur region are already among the lowest inthe country over the entire year (BBS Monthly Statistical Bulletins), and further, demandfor agricultural labor plunges between planting and harvest. The resultant drastic drop inpurchasing power for Rangpur households reliant on agricultural wage employment takes(or threatens to take) consumption below subsistence.

[Figure 1 about here.]

[Figure 2 about here.]

Several puzzling stylized facts about household and institutional characteristics andcoping strategies motivate the design of our migration experiments. First, seasonal out-migration from the monga-prone districts appears to be low despite the absence of localnon-farm employment opportunities. According to the nationally representative HIES 2005data, only 5 percent of households in Monga prone districts receive domestic remittances,while 22 percent of all Bangladeshi households do. Remittances generally under-predictout-migration rates, but the size and direction of this gap (lower migration out of thepoorest district) are still puzzling. It is more common for agricultural laborers in otherregions to migrate in search of higher wages and employment opportunities, and this isknown to be one primary mechanism by which households diversify income sources inIndia (Banerjee and Duflo 2006).

Second, inter-regional variation in income and poverty between Rangpur and the rest ofthe Bangladesh have been shown to be much larger than the inter-seasonal variation withinRangpur (Khandker 2010). This suggests smoothing strategies that take advantage of inter-regional variation (i.e. migration) rather than inter-seasonal variation (e.g. savings, credit)may hold greater promise. Moreover, an in-depth case-study of the Monga phenomenon(Zug 2006) explicitly notes that there are off-farm employment opportunities in rickshaw-pulling and construction in nearby urban areas during the monga season. To be sure,Zug 2006 points out that this is a risky proposition for many, as labor demand and wagesdrop all over rice-growing Bangladesh during that season. However, this seasonality is lesspronounced than that observed in Rangpur (Khandker 2010).

Finally, both government and large NGO monga-mitigation efforts have concentrated ondirect subsidy programs like free or highly-subsidized grain distribution (e.g. “Vulnerable

4Amartya Sen (1981) notes these price spikes and wage plunges as important causes of the 1974 faminein Bangladesh, and that the greater Rangpur districts were among the most severely affected by this famine.

5

Group Feeding,”), or food-for-work and targeted microcredit programs. These programsare expensive, and the stringent micro-credit repayment schedule may itself keep house-holds from engaging in profitable migration (Shonchoy 2010). There are structural reasonsassociated with rice production seasonality for the seasonal unemployment in Rangpur,and thus encouraging seasonal migration towards where jobs are appears to be a sensiblecomplementary policy to experiment with.

3 Design of Interventions and Experiment

The two districts where the project is conducted (Lalmonirhat and Kurigram) representthe agro-ecological zones that regularly witness the monga famine (World Bank Umar cite).We randomly selected 100 villages in these two districts and first conducted a village censusin each location in June 2008. Next we randomly selected 19 households in each villagefrom the set of households that reported (a) that they owned less than 50 decimals of land,and (b) that a household member was forced to miss meals during the prior (2007) mongaseason. We conducted a baseline survey of these 1900 households during the pre-mongaseason in July 2008.

In August 2008 the researchers randomly allocated the 100 villages into four groups:Cash, Credit, Information and Control using a pure random number generator in Stata.These treatments were subsequently implemented in collaboration with PKSF5 and theirNGO partner organizations with field presence in the two chosen districts. We trainedthe NGOs on the implementation procedure in August 2008, and they implemented theinterventions during the 2008 Monga season starting in September. 16 of the 100 studyvillages (consisting of 304 sample households) were randomly assigned to form a controlgroup. A further 16 villages (consisting of another 304 sample households) were placedin a job information only treatment. These households were given information on typesof jobs available in four pre-selected destinations, the likelihood of getting such a job andapproximate wages associated with each type of job and destination. 703 households in 37randomly selected villages were offered cash of Taka 600 (∼US$8.50) at the origin condi-tional on migration, and an additional bonus of Taka 200 (∼US$3) if the migrant reportedto us at the destination during a specified time period. We also provided exactly the sameinformation about jobs and wages to this group as in the information-only treatment. Taka600 covers a little more than the average round-trip cost of safe travel from the two origindistricts to the 4 nearby towns on which we provided job and wage information. The 589households in the final set of 31 villages were offered the same information and the sameTk 600 + Tk 200 incentive to migrate, but in the form of a zero-interest loan to be paidback at the end of the monga season (in December) rather than a cash grant.

5PKSF (Palli Karma Sahayak Foundation) is an apex micro-credit funding and capacity building orga-nizations in Bangladesh. It is a company not for profit set up by the Government of Bangladesh in 1990,for poverty alleviation through the provision of micro-credit through its Partner Organisations (POs).

6

In the 68 villages where we provided monetary incentives for people to seasonally out-migrate (37 cash + 31 credit villages), we sometimes randomly assigned additional condi-tionalities to different households within the village. For example, 12 of the 19 householdswere required to migrate in groups, and specific group members were assigned in half thecase. Also, about half the households were required to migrate to a specific destination.We will not directly analyze the effects of such conditionalities in this paper, but theseconditionalities were helpful in creating random within-village variation in specific house-holds propensity to migrate. This random variation will be useful for some of the empiricalanalysis presented later. Figure 8 provides an overview of the randomized interventionconditions.

4 Returns to Seasonal Migration

In this section we report simple “program evaluation” results on take-up of the treatment,the effects of seasonal migration on the households consumption and welfare at the origin,income and savings of the migrant at the destination, and the propensity to re-migrate in2009 after incentives are removed. After establishing the large, positive returns to seasonalmigration, in subsequent sections we turn to the question of why households were notengaging in this profitable behavior to begin with.

[Table 1 about here.]

Table 1 examines program take-ups for two incentivized groups – cash and credit puttogether, and out-migration rates between incentivized groups and not-incentivized groups– control and information put together. Against a migration rate of 58% for the incentivizedgroup (after six months of incentives offered), the migration rate for the not-incentivizedgroup stood at 36%. The statistical analysis conducted in the various panels of Table 1show that the effect of the cash and credit incentives are highly statistically significant andthat providing information has essentially a zero effect on migration propensity. There isalso no difference between providing cash and providing credit – households appear to reactvery similarly to either incentive, we therefore combine the impact of these two treatmentsfor much of our analysis. We return to provide a positive explanation of this fact in Section5.

Table 2 shows the effects of migration on consumption expenditures and other expendi-tures amongst remaining household members during the monga season. Consumption is anideal measure of the benefit of migration, aggregating as it does the impact on the familyof migrating. Using such a measure is important as it overcomes the problems associatedwith measuring the true costs and benefits of technology adoption that are highlighted,for example, in Foster and Rosenzweig (2010). The effects are calculated from a regressionwhere the choice to migrate is instrumented with whether or not a household was randomly

7

placed in the incentive group. These estimates therefore show the impact of migration onthose households that were induced to migrate by our intervention (that is the LATE).

[Table 2 about here.]

Migration of a household member during the monga season has substantial impact onthe remaining household members well-being. Compared to non-migrant households, percapita food, non-food, and caloric intake among (induced) migrant households increaseby 30% to 35%, and their monthly consumption expenditure increase by at least $4 percapita ($15 per household) due to induced migration. There are changes towards higherquality diets as food consumption shifts towards meat and child education expendituresincrease among migrant households. There is also a statistically significant increase in non-food expenditures on clothing and shoes, and expenditures of health for male members ofhouseholds.

[Table 3 about here.]

[Table 4 about here.]

The positive consumption effects on the remaining household members come from re-mittance that migrant members send from their earnings and saving. Table 3 (which shouldbe compared to Table 4) breaks down the effects. The incentivized migrants earn $110 onaverage during the lean season and save about half of that, and the average savings plusremittance is about a dollar a day. Non-incentivized migrants earn more per episode, saveand remit more per day, which indicates our induced migrants earning potentials were lessthan the average. However, on an incentive of about $6-$8, the induced migrants earn $110on average during the lean season and save and remit more than half of that, which is avery high rate of return on per dollar incentive.However, not all migrants were successfulin finding jobs at destinations, earnings and remitting to their households at origin. About20% of the migrants were unsuccessful which had consequence on remigration possibilitiesin subsequent year.

Our one-time incentive to migrate was offered in 2008 prior to the lean season. As shownin Table 1, in 2009 47% of the incentivized groups re-migrated even after the incentive wastaken away. The comparable figure for the non-incentivized group was 37% that showsalmost no change on year-to-year basis in migration rate in the control and informationgroups. We find that if a household was induced to migrate in 2008, it doubles the chances(35-45 percentage point effect) that it will migrate in 2009.

5 A Model of Risky Experimentation

In the previous section we documented three facts regarding our intervention. First, a largenumber of households were motivated to migrate in response to a relatively small incentive

8

(600 Taka). Second, there was a positive average return to migration, indicating thathouseholds were not migrating despite a positive expected profit. Third, a large portion ofthe households that were incentivized to migrate in year one continued to send a seasonalmigrant in year two and continued to earn a high return to this activity. We consider theseto be important observations, they document a situation in which a small transfer has ameaningful and lasting impact on the behavior and well being of poor households.

In this section we propose a (very) simple model of a poverty trap, and argue that itcan explain the three main findings from the previous section. The basic model is a muchstripped down version of the classic models of Kihlstrom and Laffont (1979) and Banerjeeand Newman (1991) and can be easily extended along the lines of Banerjee (2004) toencompass a more traditional view of poverty traps. To this extent we see our resultsas lending support to the basic proposition that poverty traps can exist. We then usethe model to determine additional predictions that should hold in the data. In section7 we discuss alternative theories that can also generate similar facts and argue that ourinterpretation is to be preferred on several dimensions.

Consider a household that lives for an infinite number of periods and in each perioddecides between staying at home and earning a certain income of y and sending a migrantand earning an income of y + b with probability µ(b) and y + g with probability µ(g). Weassume that g > b and we also assume that after one migration episode the householdbecomes informed about whether the return is g or b. The best way to interpret the returnin this context is that successful migration (g) requires connections at the destination.Before the first migration episode the household is unsure whether it will be able to find aconnection, however, once the connection is in place it is permanent.

We assume that the household is a discounted expected utility maximizer with utilityfunction u (that obeys all the usual assumptions), discount factor δ and that b and g aresuch that:

u(y + g) > u(y); andu(y + b) < u(y).

Given these assumptions the return to migrating is

V m =µ(g)u(y + g) + µ(b)((1− δ)u(y + b) + δu(y))

1− δ,

while the return to staying at home is

V h =u(y)1− δ

.

Figure 3 shows a pretty typical plot of V m and V h as a function of baseline income y whereV m is in pink and V h in blue. While the two curves need not cross, if they do it is usually

9

true that they cross only once with V m being optimal above some threshold income y∗.6

It is much more likely that the curves will cross like this if the outcome b is sufficiently badso that u(y + b) is in the very steep part of the utility function. In the example used togenerate Figure 3 this requires that y + b be close to 0, the point at which the derivativeof the ln utility function becomes infinite.

Given that the conditions for a single crossing are met, the model generates a povertytrap. Households with income y < y∗ never migrate, they consume their initial incomey forever. On the other hand, households that have income of y ≥ y∗ migrate and learntheir migration status. If they are able to find a connection they earn y + g forever afterand if they are not able to find a connection they return to consuming y. The long rundistribution of income is therefore as depicted in Figure 4. Thus we have a very simplepoverty trap, those with initial incomes below a certain level stagnate at that level whilethose with starting incomes above that point are able to take advantage of migration andpush their long run income up by g.

[Figure 3 about here.]

[Figure 4 about here.]

Of course, in reality incomes vary across years and those that are close in averageincome to the cutoff y∗ will eventually jump from the low income equilibrium to the highincome equilibrium. Nevertheless, poverty trap models such as this one provide a role forpolicy in helping those household whose income is sufficiently low that it is unlikely thatthey will move between the equilibria. Further, if the return g is large enough, a policythat helps households make the transition can lead to a modest transformation in people’slives.

Given this background, it seems reasonable to assume that a policy maker may wishto assist households in making the jump between equilibria, and it is natural to ask howthis is best achieved. Figure 5 compares three policies. The first panel shows the baselineincome threshold y∗, the second panel considers a once off transfer, the third an incentivesimilar to our “cash” treatment and the fourth an “insurance” program that provides atransfer only if the household migrates and realizes state b. This fourth panel considersa policy that is similar to a limited liability loan and is formally similar to our “loan”treatment. Two points should be noted. First, because the transfer increases both V m

and V h it has a much more limited impact on y∗. The alternative incentive treatment ismuch more effective in encouraging migration. Second, because the poverty trap is drivenalmost entirely by the possibility of a near subsistence outcome in state b, the insurancetreatment realizes most of the benefits of the incentive policy. The insurance policy is alsomuch more cost effective; the incentive must be paid with certainty, while the insuranceonly pays out with probability µ(b). To translate this into our practical setting, the loan

6Conditions for the crossing can be found in Banerjee (2004) for example.

10

was repaid in roughly 85% of cases and the total cost of the loan treatment (excludingoperating costs) is therefore only 15% of the cost of the incentive treatment.

[Figure 5 about here.]

These two observations translate into two important implications for our experiment.First, we can quantify what it means for the impact of the incentive to be “large”. Giventhe parameters used in Figure 5 the transfer reduces y∗ from 0.275 to 0.259. On the otherhand, the incentive decreases y∗ to 0.241, that is, the same size incentive has roughly twicethe impact.7 Another way to see this is to note that an incentive of this size has the sameimpact as an income transfer of over twice the size. We, therefore, expect our incentivescheme to have a large impact relative to yearly income variation that is roughly the samesize, and this gives scope for an incentive policy to be cost effective in moving householdsbetween equilibria. Second, because the insurance policy realizes the lions share of the gainfrom the incentive policy, we predict that empirically our “cash” and “loan” treatmentswill have a similar size impact on the migration rate.

How general are these results and when would we expect them to hold? The situationconsidered in Figure 5 has a particular feature. Household income y∗ is very low, in that itfalls in the very steep part of the utility function used for the example. This implies thatthe transfer policy has a relatively large impact on V h and therefore that the effect of theincentive is large relative to the effect of the transfer. If, however, y∗ is high relative tosubsistence then this will no longer be the case. Figure 6 illustrates. The left panel showshow a transfer will affect the curves V m and V h, while the right panel illustrates the impactof an incentive. Because income in this example is relatively far from the subsistence point(which is 0 given the ln utility function used) the transfer has only a small impact onV h and as a result the incentive and transfer policies have a similar impact. In such acontext it seems unlikely that our incentive would have much of an impact as it is smallrelative to yearly variation in income and therefore most of the households that are farfrom subsistence that would be affected by our policy will already have moved to the highincome equilibrium. This reasoning combined with the fact that the overall impact of ourincentive on V m will be largest when y is small (due to concavity of the utility function)leads us to predict that our treatments will have a larger effect on those households thatare close to subsistence.

[Figure 6 about here.]

Our model is able to generate such a stark poverty trap because we assume that there islearning about the state of the world after one migration episode. We make this assumptionfor a simple reason. In our experiment a large portion of those households that wereinduced to migrate in round 1 continue to migrate in round 2 (roughly 50%). Put simply it

7This assumes that households are uniformly distributed with respect to income.

11

is difficult to generate this degree of remigration in a model without learning. Again, this isbecause the earnings from migration operate similar to a transfer policy and the payoff tomigration would therefore have to be large relative to the yearly variation in income. Thisseems unlikely because the incentive offered (600 Taka) is small relative to yearly variationin income and it seems unlikely that the return to that 600 Taka investment is more than100%. The inclusion of a learning effect also generates two additional implications. First,if we interpret learning as finding out if you can establish a network connection, and thenkeeping that connection we should see that the incentive has most of its impact on thosethat do not already have network connections at a destination. Second, we should seelocation specific learning. That is, households that migrate to a particular location (forexogenous reasons) should continue to migrate to that location.

We plan to test the four empirical predictions of the model in the next section, so wecollect them here for reference:

Predictions 1. If our three main results are driven by a poverty trap model as presentedin this section then:

1. Cash and Credit treatments should have a similar impact on migration rates;

2. The incentive should have a larger impact on households that are close to subsistence;

3. The incentive should have a larger impact on households that do not have networkconnections at the destination; and

4. Households should exhibit location specific learning.

The model and these four predictions also help us to understand the type of situationin which we would expect incentive and insurance policies to lead to the sorts of longterm effects observed in our experiment. We should look for situations in which there ishousehold specific learning, where households are close to subsistence and where there isa reasonable chance of a negative outcome from investment. Prediction 2 also provides ananswer to the puzzle which motivated this work – i.e. why does Rangpur have such a lowmigration rate relative to the rest of Bangladesh? Our model suggests that this is becausehouseholds in Rangpur are close to subsistence at the time when it makes most sense tomigrate.

6 Additional Tests of the Model

In this section we provide evidence that is consistent with the for predictions from ourmodel.

The first prediction of our model has already been verified. The migration rate in thecash treatment is 58.96% while in the credit treatment it is 56.8% and the difference isnot statistically significant. This is consistent with the claim that the cash and credit

12

treatments should have similar impacts on the migration rate, with the credit treatmentbeing slightly less effective.

The second prediction of the model is that those who are close to subsistence shouldbe less likely to migrate and that the incentive should have the greatest impact on thosehouseholds that are close to subsistence. We use the portion of household expenditurethat goes toward food as a measure of how close the household is to subsistence. Figure7 provides kernel density plots of the distribution of this index among households thatmigrate and those that do not migrate. The left hand panel shows the plot among thecontrol villages. It is clear in these diagrams that households for that are close to subsistenceare far less likely to migrate. The right panel shows the same plot amongst the treatmentvillages. In these villages there is no such discrepancy – essentially the migrant and non-migrant households look identical on this measure in the treatment villages. These twoplots are, therefore, consistent with the model.

[Figure 7 about here.]

The third implication of our model is that the incentive should have a larger impact onthose households that do not have connections at the destination and therefore have moreto learn – or equivalently, have not yet learned whether they are g or b households. Table 5provides evidence on this conjecture. The table breaks migration episodes within the firstround data down with the first migration episode presumably being the one most affectedby our incentive. The table shows that the incentive led to an increase in the number ofhouseholds migrating that do not know someone at the destintion and that do not alreadyhave a job lead at the destination. These observations are again consistent with our model.

[Table 5 about here.]

Finally, our model predicts learning. We make two predictions here. First, householdsthat are successful in their first year should be more liklely to migrate in the second yearand this effect should be more pronounced among the incentivized migrants. Second, oneof our treatment requires households to send their migrant to a particular location. Ourmodel predicts that remigration should be disproportionately back to this same location.

Table 6 provides evidence in favor of the first question. Among the incentivized villages,households that remigrate are those that were successful in the first round – i.e. those thathave higher savings and earnings in the first round. In contrast there is a much weakerdifferent in the impact of success among non-treatment villages. This is consistent withthe model and the idea that the treatment households have something to learn, while thenon-treatment villages do not.

[Table 6 about here.]

13

Table 7 provides evidence on the second conjecture. It shows results of a multinomiallogit and confirms that households that were successful in a particular location were likelyto return to that location.

[Table 7 about here.]

7 Alternatives

In this section we discuss alternative explanations for our empirical results. The mostplausible alternative explanation is that, during the lean season, households face a strongliquidity constraint – they simply do not have the 600 Taka required to pay for a bus ticketand therefore cannot migrate. Relieving the liquidity constraint in one year could lead toongoing migration either through learning as discussed above or because increased incomerelaxes the liquidity constraint again the following year.8 Such an explanation can accountfor the majority of our facts. A large portion of the population may be liquidity constrainedin any given year leading to the large effect of the incentive, the return to migration canbe positive and yet people do not migrate because they lack can, the impact of cash andcredit treatments will be the same and, to the extent that subsistence is correlated withthe liquidity constraint, we would expect the incentive to have a greater impact on thoseclose to subsistence.

Nevertheless we do not thing a liquidity constraint can explain our results. Given thestrong repeat impact on the second year, and the small size of the transfer, a liquidityconstraint story would not hold over time. That is, we would expect all households tohave received a sufficiently strong positive income shock in the past to have overcomethe liquidity constraint and moved to the rich equilibrium prior to our intervention. Forexample, year on year average absolute deviation in weekly expenditure in our sample is307 Taka between rounds one and two and 368 Taka between rounds two and three. Thestandard deviation of absolute deviation in consumption is 635 and 508 Taka respectively.While part of this is no doubt driven by measurement error, these figures show what seemsobvious, season on season variation in incomes likely swamps the 600 Taka of our incentive.This criticism applies to all mechanisms that treat the incentive as a one time transfer.Further to this basic criticism, the lean season in this region lasts for around 3 months. Itis hard to see how any household that has sufficient funds, either as storage or recurringwages, to survive this period would be unable to find 600 Taka. This may be a very costlychoice, but it will be possible. Liquidity constraints therefore seem to be counter factual.

The above discussion applies equally to any mechanism that does not emphasize theincentive effect of our treatments. This observation, however, raises a perhaps troublingpoint for our preferred model. The basic argument against 600 Taka having a large effect

8It should be noted that some form of credit constraint is required for the poverty trap model above towork. The liquidity constraint discussed here is a more severe form of credit constraint.

14

in the form of a transfer is that yearly variation in income is larger. Therefore, for ourincentive to work it must be the case that many households would prefer to keep a onceof transfer of 600 Taka rather than spend it one migration. But this implies that the wel-fare impact of our intervention is bounded above by the benefit of giving each household600 Taka to use as they please and the households not migrating. This point is of courseobvious in a fully rational model because households always prefer to have money withoutconstraints. We leave it to the reader to determine whether they believe that it is reason-able to assume this. Alternative models that posit less rationality, either in the form ofexcessive impatience (δ low), non-rational expectations (µ(b) too low) or some reason whythe household is consistently on the poverty line despite variation in income (for examplehyperbolic disounting a la Laibson (1997) and Duflo et al. (2009)) can make some roomfor policy makers to believe that the welfare gains of this intervention are larger than thepurely rational model would imply.

8 Conclusions

We conducted a randomized experiment in which we incentivized households in Rangpur tosend a seasonal migrant to an urban area. The main results show that a small incentive ledto a large increase in the number of seasonal migrants, that the migration was successfulin the sense that it led to an average increase in consumption of around 35% and thathouseholds given the incentive in one year continued to be more likely to migrate in thefollowing year.

We argue that the results are consistent with a simple (rational) model of a poverty trapwhere households that are close to subsistence are unable to learn whether migration issuccessful due to a small possibility that migrating will turn out badly, leaving householdsconsumption below subsistence.

From a policy perspective, the results support a policy of providing micro-insurancethat mitigates the potential downside of experimentation. Our analysis also suggests thatsuch a policy will be particularly beneficial where households are close to subsistence, riskaverse and could benefit from a technology that requires individual level experimentationin order to determine profitability.

[Table 8 about here.]

References

Banerjee, A., “The Two Poverties,” in Stefan Dercon, ed., Insurance against poverty,Oxford Scholarship Online Monographs, 2004, pp. 59–76.

Banerjee, A.V. and A.F. Newman, “Risk-bearing and the theory of income distribu-tion,” The Review of Economic Studies, 1991, 58 (2), 211.

15

Duflo, E., M. Kremer, and J. Robinson, “Nudging farmers to use fertilizer: theoryand experimental evidence from Kenya,” NBER Working Paper, 2009.

Foster, A.D. and M.R. Rosenzweig, “Microeconomics of technology adoption,” Annu.Rev. Econ., 2010, 2 (1), 395–424.

Kihlstrom, R.E. and J.J. Laffont, “A general equilibrium entrepreneurial theory offirm formation based on risk aversion,” The Journal of Political Economy, 1979, 87 (4),719–748.

Laibson, David, “Golden Eggs and Hyperbolic Discounting,” Quarterly Journal of Eco-nomics, 477 1997, 113 (2), 443.

16

List of Figures

1 Seasonality of Rice Prices . . . . . . . . . . . . . . . . . . . . . . . . . . . . 182 Seasonality of Food Expenditure . . . . . . . . . . . . . . . . . . . . . . . . 193 The return to migration as a function of income . . . . . . . . . . . . . . . . 204 Long run income distribution . . . . . . . . . . . . . . . . . . . . . . . . . . 215 The effect of policy on cutoff for migration . . . . . . . . . . . . . . . . . . . 226 The effect of policy on cutoff for migration with income above subsistence . 237 Kernel Density Plots of Subsistence Index by Treatment Status . . . . . . . 24

17

Figure 1: Seasonality of Rice Prices

18

Figure 2: Seasonality of Food Expenditure

19

Figure 3: The return to migration as a function of income

6.0 6.5 7.0 7.5 8.0

5.5

6.0

6.5

7.0

u = ln, µ(g) = 0.7, b = −6, g = 3 and δ = 0.7.

20

Figure 4: Long run income distribution

5.5 6.0 6.5 7.0 7.5 8.0

4

6

7

8

9

10

11

Parameters as in Figure 3. Dashed lines represent possible outcomes and solid lines average outcomes.

21

Figure 5: The effect of policy on cutoff for migration

0.25 0.26 0.27 0.28 0.29

�4.6

�4.5

�4.4

�4.3

�4.2

�4.1

�4.0Base Case

0.25 0.26 0.27 0.28 0.29

�4.6

�4.5

�4.4

�4.3

�4.2

�4.1

�4.0Transfer

0.25 0.26 0.27 0.28 0.29

�4.6

�4.5

�4.4

�4.3

�4.2

�4.1

�4.0Incentive

0.25 0.26 0.27 0.28 0.29

�4.6

�4.5

�4.4

�4.3

�4.2

�4.1

�4.0Insurance

u = ln, µ(g) = 0.7, b = −0.2, g = 0.05 and δ = 0.7.

22

Figure 6: The effect of policy on cutoff for migration with income above subsistence

6.0 6.5 7.0 7.5 8.0

5.5

6.0

6.5

7.0

7.5Transfer

6.0 6.5 7.0 7.5 8.0

5.5

6.0

6.5

7.0

7.5Incentive

u = ln, µ(g) = 0.7, b = −6, g = 3 and δ = 0.7.

23

Figure 7: Kernel Density Plots of Subsistence Index by Treatment Status

24

List of Tables

1 Program Takeup . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 262 Effects of Migration on Expenditure Amongst Remaining Household Members 273 Earnings & Saving from Migration (Round 2) . . . . . . . . . . . . . . . . . 284 Earnings for non-migrants . . . . . . . . . . . . . . . . . . . . . . . . . . . . 295 Who Migrates by Treatment Group . . . . . . . . . . . . . . . . . . . . . . . 306 Remigration by Earnings in First Migration Episode . . . . . . . . . . . . . 317 Remigration by Earnings in First Migration Episode . . . . . . . . . . . . . 328 Structure of the Randomization . . . . . . . . . . . . . . . . . . . . . . . . . 33

25

Table 1: Program Takeup!"#$%&'(&)*+,*"-&!".%/01&"23&45,*"65+2&7"6%8&

&& &9::%*&

;<<%16%3&& =%16&4+2%>& &

45,*"65+2&7"6%&

?"8@& ! A'BCCD& & ECBFGD& & HCBIGD&

?*%356& ! HFBICD& & JEBF'D& & HGBCKD&

L2:+& ! JHB'ED& & B& & JHBIJD&

?+26*+$& ! B& & B& & JHBIAD&

&& & L2<%265M5N%3& &O+6&L2<%265M5N%3& & )/P"$0%&

! HCD& & JGD& &45,*"65+2&7"6%& ! QKBK'ER& & QKBK'IGR& &

KBKK&

! EAD& & JAD& &7%-5,*"65+2&7"6%& !! Q&KBK'ER& && QBKFK&R& &&

KBKK&

9::%*&"<<%16%3&*%:%*8&6+&@+08%@+$3&6@"6&"<<%16%3&6@%&+::%*&"23&*%<%5M%3&6@%&-+2%>B&S+-%&+:&6@%&@+08%@+$3&*%60*2%3&6@%&-+2%>&,5M%2B&=%16&4+2%>&*%:%*8&6+&6@%&5235M530"$8&T@+&"<<%16%3&6@%&+::%*&"23&.%16&6@%&-+2%>&:*+-&6@%&6*%"6-%26B&!@%&)/M"$0%&58&+#6"52%3&:*+-&6@%&35::%*%2<%&#%6T%%2&-5,*"65+2&*"6%8&+:&52<%265M5N%3&"23&2+2&52<%265M5N%3&@+08%@+$38U&6@58&*%:%*8&6+&"$$&@+08%@+$3&52&?"8@U&?*%356U&L2:+&"23&?+26*+$U&T@%6@%*&6@%>&.%16&6@%&-+2%>&+*&2+6B&

!

26

Table 2: Effects of Migration on Expenditure Amongst Remaining Household Members

! ! ! ! ! ! ! ! ! ! !! "#$%&'()*'+)!,-.-! ! "#$%&'()*'+)!,-.-!/!+01)*02(! ! !

!!

34"! !! 56! ! 34"! ! 56! !

7891!0:!;8<81&81)!69*'9$28!

! =>-?@AAA! ! BBC-DA! ! >B-CEAAA! ! BBE-FA! !,00&!.G<81&')#*8(!!

! H?D-EDI! ! H?BC-BI! ! H?D-BDI! ! H??=->I! !=B>-B!

! C@-ECAAA! ! ???-JAA! ! JJ->?AAA! ! ???-BAA! !K01!,00&!.G<81&')#*8(!

! HD-CCDI! ! HC>-JCI! ! HD-FDDI! ! HCC-C?I! !B=C-C!

! ?BC-JAAA! ! FF=-JAA! ! ?C=-CAAA! ! FFB-JAA! !L0)92!.G<81&')#*8(!

! HBB-F@I! ! H?JC-?I! ! HBB-B@I! ! H?CF-EI! !?EEF-?!

! BF?-FAAA! ! =B>-CAAA! ! B@B-=AAA! ! @@@-@AAA! !L0)92!M920*'+!'1)9N8!

! HCE-@?I! ! HBFD-?I! ! HCB-@>I! ! HBFC-DI! !BE>?-F!

! C-?@DAAA! ! ?E-J?! ! C-DD>AAA! ! ?E-E=! !L0)92!M920*'8(!:*0O!P*0)8'1!

! H?-?=@I! ! HD-=E>I! ! H?-B=@I! ! HD-BF=I! !C@->!

! C-@EJ! ! B>-BJ! ! =-E?FA! ! FF-@EA! !.G<81&')#*8(!01!789)!P*0&#+)(!

! HF-=E=I! ! H?D-?JI! ! HF-DJBI! ! H?D-B>I! !BD-B!

! ?-@C! ! %E-J>J! ! E->E>! ! %E->FD! !.G<81&')#*8(!!01!7'2N!Q!.RR(!

! H?-CJFI! ! HD-E@=I! ! H?-CF?I! ! H=->=JI! !?F-?!

! E-J@C! ! %?B-==! ! ?-@?J! ! %D-@>F! !.G<81&')#*8(!01!,'(S!

! H@-JJBI! ! HCD-?BI! ! H=-?FFI! ! HCF-EFI! !=C-C!

! %F-B?@! ! BD-?EAA! ! %C-ED>A! ! B?-EJA! !.G<81&')#*8!01!MS'2&*81T(!.&#+9)'01! ! HB-FCI! ! H?B-DDI! ! HB-CFCI! ! H?B-E@I! !

?D-B!

! @-BCFAAA! ! ?B-=B! ! @->BDAAA! ! ?B-JCA! !.G<81&')#*8(!01!M20)S'1R!91&!"S08(! ! H?-JFI! ! HD-?E?I! ! H?-JBDI! ! H=-F>JI! !

FD-D!

! ?-J@J! ! JF-F=! ! E-BEC! ! JE-DE! !.G<81&')#*8(!01!U892)S!:0*!,8O928! ! H@->FJI! ! HF>-F@I! ! H=-F?CI! ! HF>-=DI! !

@?-J!

! ?D-FBAA! ! BC-CD! ! ?>-E>AA! ! ?@-CB! !.G<81&')#*8(!0:!U892)S!:0*!7928! ! HD-?>>I! ! HFJ-BDI! ! HD-@CEI! ! HF=-=EI! !

=?-C!

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

!

!!

27

Table 3: Earnings & Saving from Migration (Round 2)

! ! ! ! ! ! ! ! !!! ! "##!$%&'()*+! ! ,)-.)*%/%0.1! ! 23*!,)-.)*%/%0.1! ! 45+!

63*(#!7(/%)&+!58!93:+.93#1!! ! ;<=>?@! ! ;@>A?A! ! ;<;<?=! ! =@B!

63*(#!C(')%)&+!58!93:+.93#1!! ! DDDD?E! ! D<@B?;! ! FF=<?<G! ! =@E!

! ! ! ! ! ! ! ! !7(/%)&+!H.'!1(8!! ! @A?F! ! @A?@! ! @D?F! ! =>@!

C(')%)&+!H.'!1(8!! ! ==?<! ! =A?B! ! BBB?@GGG! ! =EA!

I.J%**()-.+!H.'!1(8!! ! BD?F! ! BA?E! ! E;?;GGG! ! =EA!

! ! ! ! ! ! ! ! !

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

!

28

Table 4: Earnings for non-migrants

!"#$%&'' (")*'+%,)$*&-'' '

+%,)$*&-'.'/"&%,)$*&-'

0$1'2*,&3'456)*' ' 789:' ' ;:97'

0$1'2*,&3'<5)5=*' ' >897' ' >?9>'

@$"'AB=6#C)2C=5)'DCE6"&EE'456)*'F=$G62E' '>H9H'

'9'

I0$1'J*,&'456)*I'=&G&=E'2$'&5="6"BE'G=$%'K$1E'2L52'5=&',56-'$"'-5*'2$'-5*'15E6E9'<5)5=*'=&G&=E'2$'&5="6"BE'G=$%'K$1E'2L52'5=&',56-'$"'5'%$"2L'2$'%$"2L'15E6E9'M$='NL6#L'N&'-6O6-&-'1*'P?'-5*E'2$'B&2'-56)*'&5="6"BE9'@$"'AB=6#C)2C=5)'DCE6"&EEQ'5=&'&"2&=,=6E&E'$G'6"-6O6-C5)E'NL$'5=&'"$2'-&-6#52&-'2$'2L&'5B=6#C)2C=5)'1CE6"&EE9'JL&*'=&,$=2'%$"2L)*',=$G62EQ'L&"#&'N&'-6O6-&-'2L6E'"C%1&='$O&='P?'-5*E'2$'#5)#C)52&'-56)*',=$G62E9'/"&%,)$*&-'6E'-&G6"&-'5E',&$,)&'NL$'N&=&'"$2'N$=R6"B'5"-')$$R6"B'G$='5'K$1Q'$='E2$,')$$R6"B'G$='5'K$1'1&#5CE&'2L&*'#$C)-'"$2'G6"-'5"*'N$=R'5O56)51)&9'/"&%,)$*%&"2'6E'51$C2'H?S9'JL&'T5E2'#$)C%"'6E'2L&'5O&=5B&'&5="6"BE'5#=$EE'&%,)$*&-'5"-'C"&%,)$*&-'N$=R&=E9''

!

29

Table 5: Who Migrates by Treatment Group

!"#$"%&$ '(")*"#$"%&$ +*, -./)011(1

2*13.)04*3(/$) 567 897 :;<6=== :;:5

-$#("/))04*3(/$) 8:7 6>7 :;<>== :;:8

?@*1/)04*3(/$) 8A7 A>7 :;<5 :;:B

2(C1.@)04*3(/$) A87 AA7 :;:8 :;<<

!"#$"%&$ '(")*"#$"%&$ +*, -./)011(1

2*13.)04*3(/$) >67 557 :;<6=== :;:D

-$#("/))04*3(/$) >B7 567 :;<A== :;:8

?@*1/)04*3(/$) D87 957 :;<A==) :;:B

2(C1.@)04*3(/$) 9D7 9B7 :;:8 :;<9

?EFG$H)I$1#$".EJ$)(K)I$(4G$).@E.)L"(M)-(N$("$)E.)+$3%"E%("

===)4O:;:<P)==)4O:;:9P)=)4O:;<

?EFG$H)I$1#$".EJ$)(K)I$(4G$).@E.)@E/)E)Q(F)R$E/)E.)+$3%"E%("

===)4O:;:<P)==)4O:;:9P)=)4O:;<

30

Table 6: Remigration by Earnings in First Migration Episode

!"#$%"#&'() *+(,%"#&'()

-+)'.,!'/"(#0,1&+2,2"#&'3+(,4(56(3/"768 9:;9$< =>?@$A BBB

*+),4(56(3/"768 9>@?$A 9=:C$9-+)'.,D'&("(#0,1&+2,2"#&'3+(,

4(56(3/"768 ;9AC$= <:C;$9 BBB*+),4(56(3/"768 ;;@:$> ;>9;$C B

E+F06G+.8,!'/"(#0,H6&,8'I,4(56(3/"768 >@$: <@$; B

*+),4(56(3/"768 <?$> <<$:E+F06G+.8,D'&("(#0,H6&,8'I,

4(56(3/"768 @A@$: ;>$C*+),4(56(3/"768 @@>$C @A<$9 BBB

%"#&'3+(,"(,&+F(8,9

BBB,HJA$A@K,BB,HJA$A<K,B,HJA$@K,-G"0,"0,1+&,)G6,)60),)G'),5+2H'&60,)G6,)L+,G+&"7+()'.,(F2M6&0,"(,)G6,&60H653/6,."(60$

-'M.6N,D'&("(#0,MI,O62"#&'3+(

D'&("(#0,1&+2,O+F(8,=

31

Table 7: Remigration by Earnings in First Migration Episode

!"##$%&"' ()%"##$%&"' !"##$%&"' ()%"##$%%&"'*

+,-.- /0123 /04/5 /0//6 7/0/18

9:;<- /04=3 /0444 /0/81 /0//6

>-);-?' /0131 /01=6 /0/4= /0/4=

@")%,?;:)A /042B /0/24 /0//4 /0/4/

C:D?''- /013/ /0/BB /0/== /0/23

!"##$%%*?%*E$F)$E*-%*-*G,$*?)E?H?E"-'I%*D?;<-)G*J$<#$JK:)*:&*,?%*$LJ$#G$E*$-<)?);%0

MN$#G%*:&*;:?);*G:*OOO MN$#G%*:&*,-H?);*:)$*

+$%K)-K:)

32

Table 8: Structure of the Randomization

Group nature:

A. Individual

B

. Assigned group

C. S

elf-formed group

Total

Group size:

Tw

o Three

Two

Three

Destination type:

Assigned

Chosen

A

ssigned C

hosen A

ssigned C

hosen A

ssigned C

hosen A

ssigned C

hosen

Incentives:

a) Information only

304

b) (a) + conditional cash transfer 133

126

66 48

54 54

66 48

60 48

703

c) (a)+ conditional credit 105

112

42 54

42 48

42 54

36 54

589

Control group

304

Total # of households 238

238

108 102

96 102

108 102

96 102

1900

!

33